Microsoft Releases Microsoft-Decision-1, a Judgment Model That Writes No Text - $0.042 per Million Input Tokens, Output Free

On October 9, 2026, Microsoft released Microsoft-Decision-1, a decision model that returns a probability for each of a fixed set of options, on Foundry and OpenRouter. It is post-trained from Qwen3.5-9B and takes three question types: noul, choice and score. It costs $0.042 per million input tokens with no output charge, and accepts text-only input of up to 32K tokens.

Microsoft Azure API pricing decision-models classification

On October 9, 2026, Microsoft announced Microsoft-Decision-1, a model that does not produce prose. Given a set of options you define, it attaches a probability to each one. The model can be used through Microsoft Foundry and OpenRouter¹. According to Microsoft, the intended jobs are sorting and routing requests, setting priorities, checking conditions and steering workflows.

Input is billed at $0.042 per million tokens, and output tokens carry no charge. The model starts from Alibaba’s open-weight Qwen3.5-9B, which Microsoft then post-trained. Microsoft plans to rebuild later versions on top of Microsoft AI (MAI) and OpenAI models.

Decision models, which return probabilities instead of text, have been announced one after another over the past month. This site has covered Strands Decider 2B, released on October 1, and OpenAI’s Decisions API, which entered beta on October 6. This time the model comes from Microsoft and runs on its own cloud; it is the third of these this site has written about.

Three Question Types, No Reasons Attached

Each request pairs a state, holding the material to be judged, with at least one question that has a declared type. The state can be plain text or JSON. There are three question types³.

noul: whether a condition holds, returned as a probability between 0 and 1
choice: picks one of the options you supply, and also returns a probability for every option
score: rates the input on an ordered scale. The value is the average of the level indexes weighted by their probabilities, so on a four-level scale it lands somewhere from 0 to 3 and can sit between two levels

Several questions about the same state can go in a single request. In the documentation’s example, one support ticket gets a team assigned with choice, a severity rating with score, and a repeat-contact check with noul. Only numbers come back; the model gives no written explanation.

The type names match the choice / noul / score vocabulary used by Strands Decider 2B. OpenAI’s Decisions API calls its true-or-false type predicate⁵.

Requests go to /providers/microsoft/v1/systemone on your Foundry resource, and the model field takes the deployment name you chose rather than the model’s own name. You sign in with either an API key or Microsoft Entra ID, the latter being Microsoft’s preference. Deployments can be DataZoneStandard, in some regions, or GlobalStandard.

How It Differs From OpenAI’s Decisions API

Like OpenAI’s Decisions API, pricing covers input only. OpenAI charges $0.10 per million input tokens for gpt-6-luna⁵, so Microsoft-Decision-1’s $0.042 comes to 42% of that rate. The Foundry blog’s price table shows the same rate for both the US Datazone and the EU Datazone.

The inputs each one accepts are not the same. The Foundry catalog says Microsoft-Decision-1 handles text only, with no image, audio or video, and scores up to 32K tokens of input in one pass⁴. OpenAI’s Decisions API can judge images as well as text, so a task such as checking product photos for damage cannot move to Microsoft-Decision-1 for now.

Where the model runs also matters. Strands Decider 2B can run on your own GPU, while Microsoft-Decision-1 is offered as a model on Azure. The catalog files it under “Direct from Azure”, meaning it is bought and billed within an Azure account. For organizations already running models on Azure, that likely means less procurement work.

Reading Microsoft’s Numbers

All accuracy and speed figures come from Microsoft’s own tests. The company says that across 36 benchmarks with roughly 150,000 questions held out of training, the model had the best accuracy and was the fastest of the models it measured. On speed, it puts the model at 2.5 times the runner-up, H2O-Lightning-4B v1.1, and 35 times GPT-6 Sol¹. The list of compared models at the end of the post also includes GPT-6 Luna Decisions and Strands-Decider 2B.

The announcement was revised after publication. According to an editor’s note, Microsoft added benchmarks for Jev covering accuracy and calibration. Per TechCrunch, Jev is a model from TypeSafe AI, released on September 15, that outputs probabilities rather than text, and TypeSafe’s latest funding round brought in $870 million, valuing the company at $7.5 billion⁶. In a field where comparison sets are still being extended after launch, ranking claims should be read as a snapshot from the time of the announcement.

The figures from internal trials also call for care. For an Xbox Research project that sorted more than 10,000 free-text feedback items into themes, the Command Line post says quality was on par with GPT-6 Sol at more than 14 times the speed and one two-hundredth of the cost¹, while the Foundry blog says quality was on par with GPT-5 at 80 to 100 times the speed². Two official posts from the same day name different baselines and different multiples.

The release stage is not labeled consistently either. The Foundry blog says “public preview”, while the lifecycle field in the catalog reads “Generally available (GA)“⁴. Organizations whose SLA or support terms depend on release stage should check before signing up.

The Foundry blog itself recommends measuring accuracy, latency and cost per decision on representative inputs from your own workload, and checking calibration on your own data if confidence values will drive thresholds.

Where It Fits, and the Limits the Docs List

The model is aimed at high-volume work that makes bounded judgments about the same input. Examples given include triaging support tickets, prioritizing incidents, filtering content, grading model output, and choosing which model or tool should handle a request. For writing, summarizing or explaining, the documentation points to a generative LLM, and it describes putting Microsoft-Decision-1 in front as a router or gate³. Inside an agent, places where an LLM is currently asked whether to allow the next step or a tool call look like candidates to test it on.

The documentation is also clear about limits. Scores shift with how questions and options are worded and ordered, and a badly posed question still returns a number. Calibration holds up best on familiar kinds of tasks. Where the model should not be pushed into picking an answer, it suggests offering a way out, for example a “cannot tell” option, and testing whether shuffling the option order changes the result. When the outcome seriously affects a person, as with lending, hiring, housing, medical care or legal questions, it asks that the model only assist people who genuinely review the result, and never decide alone³.

If you currently build classification or routing by having an LLM return a single field through function calling or JSON output, you can now compare three decision models on the same labeled data. A rough split for now would be Microsoft-Decision-1 for text-only inputs on Azure, OpenAI if images also need judging, and Strands Decider 2B if data cannot leave your environment. OpenAI expects its Decisions API to reach GA within weeks, and whether pricing or supported models change at that point will also affect the comparison.